makalo / siamese-rpn-tensorflow Goto Github PK

View Code? Open in Web Editor NEW

127.0 5.0 29.0 24.94 MB

An re-implementation for Siamese-RPN with Tensorflow

License: MIT License

Python 100.00%

tensorflow python tracing rpn

siamese-rpn-tensorflow's Introduction

Siamese-RPN-tensorflow

Code for reproducing the results in the following paper:

High Performance Visual Tracking with Siamese Region Proposal Network
Pytorch version has been available by my classmates.
Another version zkisthebest/Siamese-RPN have lots of bugs.So I have to re-implementation by myself

Environment

python=3.6
tensorflow=1.10
cuda=9.0

Downloading VOT2013 Data

Enter http://data.votchallenge.net/vot2013/vot2013.zip in your browser
Unzip the file and move to ./data

Downloading YouTube-bb Data

git clone https://github.com/mbuckler/youtube-bb.git
python3 download.py ./data 12

Downloading ILSVRC 2015-VID Data

wget http://bvisionweb1.cs.unc.edu/ilsvrc2015/ILSVRC2015_VID.tar.gz

Performance

The red box is for tracing

Visualization for debug

bbox in detection

red -- the groundtruth
black -- bbox with highest score
other colors -- bbox with scores from second to tenth.

Training and Evaluation

If your data format is the same as VOT 2013, you can run the code directly. If not, you need to change the utils/image_reader.py or convert the data format to VOT format.

To train Siamese-RPN:

python train.py

If you want to see if the training is reasonable in the course of training, you can choose to turn on debug.Just change the init() in train.py

self.is_debug=True

This will result in a debug folder where you can see pictures of the training process, with groundtruth in red and box in top 10 scores in other colors.

To test Siamese-RPN:

To test series of images like VOT format

If you want to test a series of images captured from the video, you need to assign new values img_pathand img_label in config.py, which are the files of your image's path and label, respectively. Then execute the following commands

python test.py

This command will automatically synthesize videos from image sequences, and also synthesize videos from processed images, which are saved in. / data / vedio

To test a vedio

If you are testing a video, you need to put the video in./data/vedio. You can run the following command and select the object you want to track in the first frame according to the program prompt at the beginning.

python vedio_test.py test.mp4

The 'test.mp4' is the name of your vedio

Model

I will provide the well-trained model in the next few days

siamese-rpn-tensorflow's People

Contributors

Stargazers

Watchers

siamese-rpn-tensorflow's Issues

請問model能否提供一下呢?

請問是否有您已經訓練好的model能demo您的結果嗎?

nan loss

I use your implemented loss module to my SiamesRPN++. It was working fine after some first epochs, but the loss is getting to 'nan' after that. Do you know why does it happen?
Btw, Can you implementation work with batch size is higher than 1?

step=224,loss=0.21591535210609436,cls_loss=0.22106705605983734,reg_loss=0.04982298985123634,lr=0.0010000000474974513,time=0.13663458824157715 step=225,loss=0.23805415630340576,cls_loss=0.18039944767951965,reg_loss=0.018808992579579353,lr=0.0010000000474974513,time=0.14361572265625 step=226,loss=0.22827929258346558,cls_loss=0.22993800044059753,reg_loss=0.027202464640140533,lr=0.0010000000474974513,time=0.14561033248901367 step=227,loss=0.18403905630111694,cls_loss=0.20121710002422333,reg_loss=0.019027791917324066,lr=0.0010000000474974513,time=0.1216745376586914 step=228,loss=nan,cls_loss=nan,reg_loss=nan,lr=0.0010000000474974513,time=0.17054390907287598 step=229,loss=nan,cls_loss=nan,reg_loss=nan,lr=0.0010000000474974513,time=0.12932848930358887 step=230,loss=nan,cls_loss=nan,reg_loss=nan,lr=0.0010000000474974513,time=0.12520909309387207 step=231,loss=nan,cls_loss=nan,reg_loss=nan,lr=0.0010000000474974513,time=0.13763189315795898

You might make a mistake in image_reader_cuda.py, line 78, index_d might be out of range
index_d=[tf.cond(tf.greater(index_t-interval,node_min),lambda:index_t-interval,lambda:index_t+interval),tf.cond(tf.less(index_t+interval,node_max),lambda:index_t+interval,lambda:index_t-interval)]
For example, if length of an input array is 120, index_t = 60 (random between 0 and 120), interval = 98 (random between 30 and 100) => index_d = [158, -38] would be wrong

求教？？？怎么把 ILSVRC 2015-VID转化成vot2013的文件格式？

费了好大力气把 ILSVRC 2015-VID数据下载下来，发现文件格式与vot2013不同，不能训练，衰......

特来求教作者 ILSVRC 2015-VID数据转化成vot2013格式的方法。万分感谢！

Hi @makalo, how can we use Siamese-RPN for multi-person tracking?

my bbox have some problem

I try to train this net by myself.I just used VOT and when I test ,I found the problem that all the boxes are the same size,so the accuracy is not well.It is caused by the error in code or something else?

与训练模型

作者可以提供一下训练好的模型吗

请问是如何预处理的数据

root_dir/list.txt应该不是VID数据集中的东西吧

hello,thank you for your do your implementation,do you have evaluation results on vot2016 or other mainstream datasets?

Well done! Prof Wang.

Nice work

关于网络训练

你好，请问youtube-BB全部视频下载下来有多大呢？

原论文中是读取了预训练模型的参数，微调后两层，但您的代码里好像是直接全部层的参数都自己训练了？效果怎么样呢？

The tracker is getting crazy

Hi,

I'm trying to run the demo, I select a bounding box, but instead of tracking the person, the bounding box just moves from one corner of the screen to another. Do I need to resize my video or do any other preprocessing steps?

Thanks,
Zvi

pre-trained model

Hello, I really appreciate your logical implementation of this complex network, I’ve learned a lot. But the training set is too big. I'm wondering if you could please share the pre-trained model, my email address is 17863135811@163. com, thanks a lot!